Ask a simple question in your next AI planning meeting: what is this customer's current credit limit? Then watch what happens. Someone checks the ERP. Someone else says the CRM has a newer note. A third person mentions the spreadsheet finance actually uses. Three answers, three sources, no referee.
That moment — not model selection, not prompt engineering — is why most AI pilots quietly die. An assistant that retrieves confidently from a contradictory data set produces confident contradictions. The fix isn't a better model. It's deciding, in writing, which system is allowed to be right.
A system of record is a decision, not a database
Teams treat "system of record" as a technical label that gets applied after the fact. It isn't. It's a governance decision you make on purpose: for a given entity, one system is authoritative, and every other system is a consumer that displays a copy.
The entities worth naming are usually short list: customer, contact, product, price, inventory quantity, open order, invoice, employee, project. For each one, you need an answer to four questions. Which system holds the truth. Which human role owns changes to it. How downstream systems receive updates. What happens when two systems disagree.
If you can't answer those four for an entity, you aren't ready to let AI reason over it.
Authority is usually split by field, not by system
The common mistake is declaring "the ERP is our source of truth" and calling the job done. In practice, authority splits at the field level, and that's fine — as long as the split is deliberate.
Your CRM probably should own relationship fields: owner, stage, next step, account notes. Your ERP should own financial and operational fields: credit terms, pricing, tax configuration, on-hand quantity, invoice status. Your website or e-commerce platform might own marketing content for a product while the ERP owns its cost and availability. Same customer record, same product record, different fields, different referees.
Write this down as a plain table your operators can read. Entity, field group, system of record, owning role, update path. It fits on one page, and it is the most useful artifact you can hand a team building AI workflows or integrations.
Duplicates stop being a reporting nuisance and start being a liability
Every mid-market company carries duplicates: three versions of the same account from three different sales reps, a vendor entered twice with a comma in a different place. Historically this showed up as a messy report and someone cleaned it before the board meeting.
An AI assistant doesn't clean it before the board meeting. It picks one and answers. So before a pilot, settle three things:
- Match keys — what combination of fields makes two records the same entity (tax ID, domain, account number — not company name)
- Merge authority — which role is allowed to merge, and which fields survive a merge
- Prevention — where duplicates get created today, and what check belongs at that entry point
That's the one list in this post, because the rest of this work isn't a checklist. It's a series of judgment calls about how your business actually runs.
Add an "as of" to everything
Data that's correct and stale is still wrong at the moment someone acts on it. Inventory synced nightly is fine for a planning conversation and dangerous for a customer promise. The answer isn't to make everything real-time — that costs money and adds fragility. The answer is to be honest about freshness.
For each authoritative field, record how often downstream copies refresh, and make that visible wherever the data appears. An AI answer that says "credit limit is X, as of this morning's sync" is a useful answer. One that just says "X" is a gamble.
Prove it on one question before you scale it
You don't need a full data governance program to start. Pick the single question your team re-asks every week — the one that generates Slack threads and phone calls. Can we ship this today? What did we quote them last time? Are they past due?
Make that one question answerable from one authoritative path. Name the source, fix the duplicates that affect it, set the refresh cadence, and assign the human owner. Then point your pilot at it. If the assistant gets that right consistently, you've learned something real about your data foundation. If it doesn't, you've found the gap cheaply, before it was embedded in a dozen workflows.
Where this usually lands
Most of the time, the work of assigning authority exposes a structural problem: the ERP, the CRM, the website, and the operational spreadsheets were never designed to agree. Integration patches the symptom. A unified platform — one place where the authoritative record lives and every tool reads from it — fixes the cause.
That's the work we do at Infraxio: ERP implementation, systems integration, and a Business Hub that gives your tools one version of the truth to point at. If your AI plans are stalling on data nobody trusts, reach out. We'll help you map the authority before you spend another dollar on the model.